Llama
By Meta
Llama is Meta's family of open-weight large language models, spanning multiple sizes and generations, designed to give researchers, startups, and enterprises access to strong LLM capabilities that can be run, fine-tuned, and deployed…
Definition
Llama is Meta's family of open-weight large language models, spanning multiple sizes and generations, designed to give researchers, startups, and enterprises access to strong LLM capabilities that can be run, fine-tuned, and deployed outside a closed API.
Overview
Meta released the first LLaMA models in February 2023, initially to approved researchers rather than the general public. Weights leaked publicly within days, which—combined with growing interest in open alternatives to closed models from OpenAI and Anthropic—helped kick off a wave of community fine-tuning and experimentation. Meta followed with Llama 2 in mid-2023 under a more permissive commercial license, then Llama 3 and later generations with larger context windows and improved reasoning. Like other modern LLMs, Llama models are Transformer-based and pretrained on large text corpora, then instruction-tuned and safety-aligned to follow prompts as chat assistants. Because the weights are downloadable, Llama models can be run entirely on local or private infrastructure using tools like Ollama, fine-tuned for domain-specific tasks, and served through frameworks like PyTorch or via managed endpoints on platforms such as Amazon Bedrock. Llama's openness has made it a common base model for teams building custom assistants, for researchers studying model behavior, and for frameworks like LangChain that need a self-hostable option alongside hosted APIs. It's frequently discussed alongside other open-weight families such as Mistral AI's models as an alternative to closed, API-only LLMs.
Key Features
- Open, downloadable model weights across multiple parameter sizes
- Commercially permissive license, with some usage restrictions at very large scale
- Transformer-based architecture pretrained on large, diverse text corpora
- Instruction-tuned chat variants alongside base pretrained models
- Support for fine-tuning on private or domain-specific datasets
- Runs locally via tools like Ollama or llama.cpp, or via cloud-hosted endpoints
- Growing context window and multilingual support across newer generations
- Strong ecosystem of community fine-tunes and derivative models